The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
_ „Outside the buffer 
Non-ground 
‘'The boundaries of the buffer 
using a multiple of the STD 
Figure 9: Identified outliners based on a buffer surrounding the 
estimated plane. 
3. EXPERIMENTS 
The proposed algorithm was tested using the set of artificial 
data shown in Figure 4, 5 and 6. For the simulated datasets, our 
results demonstrate 100% accurate classification of ground and 
non-ground points. The results have shown that this algorithm 
can handle the simulated sloping and hilly data effectively. This 
approach was also tested on real LiDAR data. In comparing our 
results with the ground truth, the number of misclassified points 
divided by the total number of points can give us the error rate, 
which, in this case, was calculated as 4.6896% (Chang et al., 
2007). These results have demonstrated that our approach can 
perform well with highly complex and unpredictable data from 
an urban area. 
ground and ground points from one another successfully. The 
results have also shown that the algorithm performs effectively 
in simulated hilly terrain and in urban areas. In a comparison 
with the results obtained with the TerraScan software, our 
algorithm demonstrated the capability of producing more 
competitive outputs. 
Future research will be extended to more complex scenes. In the 
next stage of research, non-ground points can also be classified 
into different objects such as buildings, trees, and cars, etc. 
Multi-return and intensity information will be taken into 
(a) (b) 
Figure 10: (a) The referenced aerial photo over the area covered 
by the LiDAR dataset, (b) The resampled DSM using LiDAR 
data. 
We also compared our results with those produced using 
TerraScan. As shown in Figure 10(a), we chose an experimental 
area around the C-Train track near the University of Calgary. 
One can see a C-Train track extending into a tunnel under the 
ground in Figure 10(a). In cases like this, the default parameters 
of our algorithm are good enough to produce acceptable results. 
The parameters for ground and non-ground classification using 
TerraScan, on the other hand, need to be adjusted iteratively. 
After fine-tuning the parameters, we computed the best results 
from TerraScan and compared them with our results. Figures 
11(a) and 11(b) show the extracted ground points and non 
ground points using the proposed approach in this paper, while 
the extracted terrain point and non-ground points using 
TerraScan are shown in Figures 12(a ) and 12(b). 
The experimental results show that our algorithm can produce 
competitive results when compared with those obtained from 
TerraScan. In some areas, our approach can delivered better 
results. The default parameters of our algorithm can produce 
stable results in most cases; however, the parameters for the 
TerraScan function need to be adjusted iteratively for each case. 
Because the function of non-ground and ground point 
classification in the TerraScan software is designed mainly for 
DTM generation, the accuracy of the ground and non-ground 
classification is not so critical for the purpose of approximated 
DTM generation. Once enough ground points can be sampled, a 
DTM can be computed using an interpolation method. 
4. CONCLUSION 
(a) (b) 
Figure 11: (a). Ground points and (b). non-ground points 
extracted using the proposed method. 
This research presented a robust algorithm for the automated 
extraction of non-ground points from LiDAR point clouds by 
detecting points that produce occlusions. Following the 
occlusion detection, a statistical filter can be used to remove the 
effects of the terrain roughness and noise. Throughout the 
experiments, the proposed procedure separated the LiDAR non- 
Figure 12: (a). Ground points and (b). non-ground points 
extracted using TerraScan.